We upgraded the AI model that decides whether a new error belongs in an existing issue or gets its own, reducing incorrect merges by 50% and duplicate issues by 20%.

What changed

The previous model used off-the-shelf code embeddings. It worked, but it wasn't trained for the nuance of stack trace similarity, which meant errors with different root causes would sometimes get merged into the same issue. When that happens you lose the signal that something new is broken, and the fixes are rarely the same.

The new model is trained on hundreds of thousands of real Sentry stack traces from consenting projects, labeled against our own internal grouping guidelines, and calibrated to err on the side of separation when a decision is ambiguous.

That means when an issue lands in your feed, it’s more likely to represent a distinct problem, with its own root cause and fix. For you, that means less time triaging noise, more time resolving real issues. You can read more about how we built it in our blog post.